virt/kvm/eventfd.c

   1 /*
   2  * kvm eventfd support - use eventfd objects to signal various KVM events
   3  *
   4  * Copyright 2009 Novell.  All Rights Reserved.
   5  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   6  *
   7  * Author:
   8  *      Gregory Haskins <ghaskins@novell.com>
   9  *
  10  * This file is free software; you can redistribute it and/or modify
  11  * it under the terms of version 2 of the GNU General Public License
  12  * as published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software Foundation,
  21  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
  22  */
  23
  24 #include <linux/kvm_host.h>
  25 #include <linux/kvm.h>
  26 #include <linux/kvm_irqfd.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/syscalls.h>
  29 #include <linux/wait.h>
  30 #include <linux/poll.h>
  31 #include <linux/file.h>
  32 #include <linux/list.h>
  33 #include <linux/eventfd.h>
  34 #include <linux/kernel.h>
  35 #include <linux/srcu.h>
  36 #include <linux/slab.h>
  37 #include <linux/seqlock.h>
  38 #include <linux/irqbypass.h>
  39 #include <trace/events/kvm.h>
  40
  41 #include <kvm/iodev.h>
  42
  43 #ifdef CONFIG_HAVE_KVM_IRQFD
  44
  45 static struct workqueue_struct *irqfd_cleanup_wq;
  46
  47 static void
  48 irqfd_inject(struct work_struct *work)
  49 {
  50         struct kvm_kernel_irqfd *irqfd =
  51                 container_of(work, struct kvm_kernel_irqfd, inject);
  52         struct kvm *kvm = irqfd->kvm;
  53
  54         if (!irqfd->resampler) {
  55                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
  56                                 false);
  57                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
  58                                 false);
  59         } else
  60                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
  61                             irqfd->gsi, 1, false);
  62 }
  63
  64 /*
  65  * Since resampler irqfds share an IRQ source ID, we de-assert once
  66  * then notify all of the resampler irqfds using this GSI.  We can't
  67  * do multiple de-asserts or we risk racing with incoming re-asserts.
  68  */
  69 static void
  70 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
  71 {
  72         struct kvm_kernel_irqfd_resampler *resampler;
  73         struct kvm *kvm;
  74         struct kvm_kernel_irqfd *irqfd;
  75         int idx;
  76
  77         resampler = container_of(kian,
  78                         struct kvm_kernel_irqfd_resampler, notifier);
  79         kvm = resampler->kvm;
  80
  81         kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
  82                     resampler->notifier.gsi, 0, false);
  83
  84         idx = srcu_read_lock(&kvm->irq_srcu);
  85
  86         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
  87                 eventfd_signal(irqfd->resamplefd, 1);
  88
  89         srcu_read_unlock(&kvm->irq_srcu, idx);
  90 }
  91
  92 static void
  93 irqfd_resampler_shutdown(struct kvm_kernel_irqfd *irqfd)
  94 {
  95         struct kvm_kernel_irqfd_resampler *resampler = irqfd->resampler;
  96         struct kvm *kvm = resampler->kvm;
  97
  98         mutex_lock(&kvm->irqfds.resampler_lock);
  99
 100         list_del_rcu(&irqfd->resampler_link);
 101         synchronize_srcu(&kvm->irq_srcu);
 102
 103         if (list_empty(&resampler->list)) {
 104                 list_del(&resampler->link);
 105                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
 106                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 107                             resampler->notifier.gsi, 0, false);
 108                 kfree(resampler);
 109         }
 110
 111         mutex_unlock(&kvm->irqfds.resampler_lock);
 112 }
 113
 114 /*
 115  * Race-free decouple logic (ordering is critical)
 116  */
 117 static void
 118 irqfd_shutdown(struct work_struct *work)
 119 {
 120         struct kvm_kernel_irqfd *irqfd =
 121                 container_of(work, struct kvm_kernel_irqfd, shutdown);
 122         u64 cnt;
 123
 124         /*
 125          * Synchronize with the wait-queue and unhook ourselves to prevent
 126          * further events.
 127          */
 128         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 129
 130         /*
 131          * We know no new events will be scheduled at this point, so block
 132          * until all previously outstanding events have completed
 133          */
 134         flush_work(&irqfd->inject);
 135
 136         if (irqfd->resampler) {
 137                 irqfd_resampler_shutdown(irqfd);
 138                 eventfd_ctx_put(irqfd->resamplefd);
 139         }
 140
 141         /*
 142          * It is now safe to release the object's resources
 143          */
 144 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 145         irq_bypass_unregister_consumer(&irqfd->consumer);
 146 #endif
 147         eventfd_ctx_put(irqfd->eventfd);
 148         kfree(irqfd);
 149 }
 150
 151
 152 /* assumes kvm->irqfds.lock is held */
 153 static bool
 154 irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
 155 {
 156         return list_empty(&irqfd->list) ? false : true;
 157 }
 158
 159 /*
 160  * Mark the irqfd as inactive and schedule it for removal
 161  *
 162  * assumes kvm->irqfds.lock is held
 163  */
 164 static void
 165 irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
 166 {
 167         BUG_ON(!irqfd_is_active(irqfd));
 168
 169         list_del_init(&irqfd->list);
 170
 171         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
 172 }
 173
 174 int __attribute__((weak)) kvm_arch_set_irq(
 175                                 struct kvm_kernel_irq_routing_entry *irq,
 176                                 struct kvm *kvm, int irq_source_id,
 177                                 int level,
 178                                 bool line_status)
 179 {
 180         return -EWOULDBLOCK;
 181 }
 182
 183 /*
 184  * Called with wqh->lock held and interrupts disabled
 185  */
 186 static int
 187 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
 188 {
 189         struct kvm_kernel_irqfd *irqfd =
 190                 container_of(wait, struct kvm_kernel_irqfd, wait);
 191         unsigned long flags = (unsigned long)key;
 192         struct kvm_kernel_irq_routing_entry irq;
 193         struct kvm *kvm = irqfd->kvm;
 194         unsigned seq;
 195         int idx;
 196
 197         if (flags & POLLIN) {
 198                 idx = srcu_read_lock(&kvm->irq_srcu);
 199                 do {
 200                         seq = read_seqcount_begin(&irqfd->irq_entry_sc);
 201                         irq = irqfd->irq_entry;
 202                 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
 203                 /* An event has been signaled, inject an interrupt */
 204                 if (irq.type == KVM_IRQ_ROUTING_MSI)
 205                         kvm_set_msi(&irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 206                                         false);
 207                 else if (kvm_arch_set_irq(&irq, kvm,
 208                                           KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 209                                           false) == -EWOULDBLOCK)
 210                         schedule_work(&irqfd->inject);
 211                 srcu_read_unlock(&kvm->irq_srcu, idx);
 212         }
 213
 214         if (flags & POLLHUP) {
 215                 /* The eventfd is closing, detach from KVM */
 216                 unsigned long flags;
 217
 218                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
 219
 220                 /*
 221                  * We must check if someone deactivated the irqfd before
 222                  * we could acquire the irqfds.lock since the item is
 223                  * deactivated from the KVM side before it is unhooked from
 224                  * the wait-queue.  If it is already deactivated, we can
 225                  * simply return knowing the other side will cleanup for us.
 226                  * We cannot race against the irqfd going away since the
 227                  * other side is required to acquire wqh->lock, which we hold
 228                  */
 229                 if (irqfd_is_active(irqfd))
 230                         irqfd_deactivate(irqfd);
 231
 232                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
 233         }
 234
 235         return 0;
 236 }
 237
 238 static void
 239 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 240                         poll_table *pt)
 241 {
 242         struct kvm_kernel_irqfd *irqfd =
 243                 container_of(pt, struct kvm_kernel_irqfd, pt);
 244         add_wait_queue(wqh, &irqfd->wait);
 245 }
 246
 247 /* Must be called under irqfds.lock */
 248 static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
 249 {
 250         struct kvm_kernel_irq_routing_entry *e;
 251         struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
 252         int n_entries;
 253
 254         n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
 255
 256         write_seqcount_begin(&irqfd->irq_entry_sc);
 257
 258         e = entries;
 259         if (n_entries == 1)
 260                 irqfd->irq_entry = *e;
 261         else
 262                 irqfd->irq_entry.type = 0;
 263
 264         write_seqcount_end(&irqfd->irq_entry_sc);
 265 }
 266
 267 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 268 void __attribute__((weak)) kvm_arch_irq_bypass_stop(
 269                                 struct irq_bypass_consumer *cons)
 270 {
 271 }
 272
 273 void __attribute__((weak)) kvm_arch_irq_bypass_start(
 274                                 struct irq_bypass_consumer *cons)
 275 {
 276 }
 277
 278 int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
 279                                 struct kvm *kvm, unsigned int host_irq,
 280                                 uint32_t guest_irq, bool set)
 281 {
 282         return 0;
 283 }
 284 #endif
 285
 286 static int
 287 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
 288 {
 289         struct kvm_kernel_irqfd *irqfd, *tmp;
 290         struct fd f;
 291         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
 292         int ret;
 293         unsigned int events;
 294         int idx;
 295
 296         if (!kvm_arch_intc_initialized(kvm))
 297                 return -EAGAIN;
 298
 299         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
 300         if (!irqfd)
 301                 return -ENOMEM;
 302
 303         irqfd->kvm = kvm;
 304         irqfd->gsi = args->gsi;
 305         INIT_LIST_HEAD(&irqfd->list);
 306         INIT_WORK(&irqfd->inject, irqfd_inject);
 307         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 308         seqcount_init(&irqfd->irq_entry_sc);
 309
 310         f = fdget(args->fd);
 311         if (!f.file) {
 312                 ret = -EBADF;
 313                 goto out;
 314         }
 315
 316         eventfd = eventfd_ctx_fileget(f.file);
 317         if (IS_ERR(eventfd)) {
 318                 ret = PTR_ERR(eventfd);
 319                 goto fail;
 320         }
 321
 322         irqfd->eventfd = eventfd;
 323
 324         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 325                 struct kvm_kernel_irqfd_resampler *resampler;
 326
 327                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 328                 if (IS_ERR(resamplefd)) {
 329                         ret = PTR_ERR(resamplefd);
 330                         goto fail;
 331                 }
 332
 333                 irqfd->resamplefd = resamplefd;
 334                 INIT_LIST_HEAD(&irqfd->resampler_link);
 335
 336                 mutex_lock(&kvm->irqfds.resampler_lock);
 337
 338                 list_for_each_entry(resampler,
 339                                     &kvm->irqfds.resampler_list, link) {
 340                         if (resampler->notifier.gsi == irqfd->gsi) {
 341                                 irqfd->resampler = resampler;
 342                                 break;
 343                         }
 344                 }
 345
 346                 if (!irqfd->resampler) {
 347                         resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
 348                         if (!resampler) {
 349                                 ret = -ENOMEM;
 350                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 351                                 goto fail;
 352                         }
 353
 354                         resampler->kvm = kvm;
 355                         INIT_LIST_HEAD(&resampler->list);
 356                         resampler->notifier.gsi = irqfd->gsi;
 357                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 358                         INIT_LIST_HEAD(&resampler->link);
 359
 360                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 361                         kvm_register_irq_ack_notifier(kvm,
 362                                                       &resampler->notifier);
 363                         irqfd->resampler = resampler;
 364                 }
 365
 366                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 367                 synchronize_srcu(&kvm->irq_srcu);
 368
 369                 mutex_unlock(&kvm->irqfds.resampler_lock);
 370         }
 371
 372         /*
 373          * Install our own custom wake-up handling so we are notified via
 374          * a callback whenever someone signals the underlying eventfd
 375          */
 376         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 377         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 378
 379         spin_lock_irq(&kvm->irqfds.lock);
 380
 381         ret = 0;
 382         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 383                 if (irqfd->eventfd != tmp->eventfd)
 384                         continue;
 385                 /* This fd is used for another irq already. */
 386                 ret = -EBUSY;
 387                 spin_unlock_irq(&kvm->irqfds.lock);
 388                 goto fail;
 389         }
 390
 391         idx = srcu_read_lock(&kvm->irq_srcu);
 392         irqfd_update(kvm, irqfd);
 393         srcu_read_unlock(&kvm->irq_srcu, idx);
 394
 395         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 396
 397         spin_unlock_irq(&kvm->irqfds.lock);
 398
 399         /*
 400          * Check if there was an event already pending on the eventfd
 401          * before we registered, and trigger it as if we didn't miss it.
 402          */
 403         events = f.file->f_op->poll(f.file, &irqfd->pt);
 404
 405         if (events & POLLIN)
 406                 schedule_work(&irqfd->inject);
 407
 408         /*
 409          * do not drop the file until the irqfd is fully initialized, otherwise
 410          * we might race against the POLLHUP
 411          */
 412         fdput(f);
 413 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 414         irqfd->consumer.token = (void *)irqfd->eventfd;
 415         irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
 416         irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
 417         irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
 418         irqfd->consumer.start = kvm_arch_irq_bypass_start;
 419         ret = irq_bypass_register_consumer(&irqfd->consumer);
 420         if (ret)
 421                 pr_info("irq bypass consumer (token %p) registration fails: %d\n",
 422                                 irqfd->consumer.token, ret);
 423 #endif
 424
 425         return 0;
 426
 427 fail:
 428         if (irqfd->resampler)
 429                 irqfd_resampler_shutdown(irqfd);
 430
 431         if (resamplefd && !IS_ERR(resamplefd))
 432                 eventfd_ctx_put(resamplefd);
 433
 434         if (eventfd && !IS_ERR(eventfd))
 435                 eventfd_ctx_put(eventfd);
 436
 437         fdput(f);
 438
 439 out:
 440         kfree(irqfd);
 441         return ret;
 442 }
 443
 444 bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
 445 {
 446         struct kvm_irq_ack_notifier *kian;
 447         int gsi, idx;
 448
 449         idx = srcu_read_lock(&kvm->irq_srcu);
 450         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 451         if (gsi != -1)
 452                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 453                                          link)
 454                         if (kian->gsi == gsi) {
 455                                 srcu_read_unlock(&kvm->irq_srcu, idx);
 456                                 return true;
 457                         }
 458
 459         srcu_read_unlock(&kvm->irq_srcu, idx);
 460
 461         return false;
 462 }
 463 EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
 464
 465 void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
 466 {
 467         struct kvm_irq_ack_notifier *kian;
 468
 469         hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 470                                  link)
 471                 if (kian->gsi == gsi)
 472                         kian->irq_acked(kian);
 473 }
 474
 475 void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
 476 {
 477         int gsi, idx;
 478
 479         trace_kvm_ack_irq(irqchip, pin);
 480
 481         idx = srcu_read_lock(&kvm->irq_srcu);
 482         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 483         if (gsi != -1)
 484                 kvm_notify_acked_gsi(kvm, gsi);
 485         srcu_read_unlock(&kvm->irq_srcu, idx);
 486 }
 487
 488 void kvm_register_irq_ack_notifier(struct kvm *kvm,
 489                                    struct kvm_irq_ack_notifier *kian)
 490 {
 491         mutex_lock(&kvm->irq_lock);
 492         hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
 493         mutex_unlock(&kvm->irq_lock);
 494         kvm_vcpu_request_scan_ioapic(kvm);
 495 }
 496
 497 void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
 498                                     struct kvm_irq_ack_notifier *kian)
 499 {
 500         mutex_lock(&kvm->irq_lock);
 501         hlist_del_init_rcu(&kian->link);
 502         mutex_unlock(&kvm->irq_lock);
 503         synchronize_srcu(&kvm->irq_srcu);
 504         kvm_vcpu_request_scan_ioapic(kvm);
 505 }
 506 #endif
 507
 508 void
 509 kvm_eventfd_init(struct kvm *kvm)
 510 {
 511 #ifdef CONFIG_HAVE_KVM_IRQFD
 512         spin_lock_init(&kvm->irqfds.lock);
 513         INIT_LIST_HEAD(&kvm->irqfds.items);
 514         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 515         mutex_init(&kvm->irqfds.resampler_lock);
 516 #endif
 517         INIT_LIST_HEAD(&kvm->ioeventfds);
 518 }
 519
 520 #ifdef CONFIG_HAVE_KVM_IRQFD
 521 /*
 522  * shutdown any irqfd's that match fd+gsi
 523  */
 524 static int
 525 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 526 {
 527         struct kvm_kernel_irqfd *irqfd, *tmp;
 528         struct eventfd_ctx *eventfd;
 529
 530         eventfd = eventfd_ctx_fdget(args->fd);
 531         if (IS_ERR(eventfd))
 532                 return PTR_ERR(eventfd);
 533
 534         spin_lock_irq(&kvm->irqfds.lock);
 535
 536         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 537                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 538                         /*
 539                          * This clearing of irq_entry.type is needed for when
 540                          * another thread calls kvm_irq_routing_update before
 541                          * we flush workqueue below (we synchronize with
 542                          * kvm_irq_routing_update using irqfds.lock).
 543                          */
 544                         write_seqcount_begin(&irqfd->irq_entry_sc);
 545                         irqfd->irq_entry.type = 0;
 546                         write_seqcount_end(&irqfd->irq_entry_sc);
 547                         irqfd_deactivate(irqfd);
 548                 }
 549         }
 550
 551         spin_unlock_irq(&kvm->irqfds.lock);
 552         eventfd_ctx_put(eventfd);
 553
 554         /*
 555          * Block until we know all outstanding shutdown jobs have completed
 556          * so that we guarantee there will not be any more interrupts on this
 557          * gsi once this deassign function returns.
 558          */
 559         flush_workqueue(irqfd_cleanup_wq);
 560
 561         return 0;
 562 }
 563
 564 int
 565 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 566 {
 567         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 568                 return -EINVAL;
 569
 570         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 571                 return kvm_irqfd_deassign(kvm, args);
 572
 573         return kvm_irqfd_assign(kvm, args);
 574 }
 575
 576 /*
 577  * This function is called as the kvm VM fd is being released. Shutdown all
 578  * irqfds that still remain open
 579  */
 580 void
 581 kvm_irqfd_release(struct kvm *kvm)
 582 {
 583         struct kvm_kernel_irqfd *irqfd, *tmp;
 584
 585         spin_lock_irq(&kvm->irqfds.lock);
 586
 587         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 588                 irqfd_deactivate(irqfd);
 589
 590         spin_unlock_irq(&kvm->irqfds.lock);
 591
 592         /*
 593          * Block until we know all outstanding shutdown jobs have completed
 594          * since we do not take a kvm* reference.
 595          */
 596         flush_workqueue(irqfd_cleanup_wq);
 597
 598 }
 599
 600 /*
 601  * Take note of a change in irq routing.
 602  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 603  */
 604 void kvm_irq_routing_update(struct kvm *kvm)
 605 {
 606         struct kvm_kernel_irqfd *irqfd;
 607
 608         spin_lock_irq(&kvm->irqfds.lock);
 609
 610         list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
 611                 irqfd_update(kvm, irqfd);
 612
 613 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 614                 if (irqfd->producer) {
 615                         int ret = kvm_arch_update_irqfd_routing(
 616                                         irqfd->kvm, irqfd->producer->irq,
 617                                         irqfd->gsi, 1);
 618                         WARN_ON(ret);
 619                 }
 620 #endif
 621         }
 622
 623         spin_unlock_irq(&kvm->irqfds.lock);
 624 }
 625
 626 /*
 627  * create a host-wide workqueue for issuing deferred shutdown requests
 628  * aggregated from all vm* instances. We need our own isolated single-thread
 629  * queue to prevent deadlock against flushing the normal work-queue.
 630  */
 631 int kvm_irqfd_init(void)
 632 {
 633         irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
 634         if (!irqfd_cleanup_wq)
 635                 return -ENOMEM;
 636
 637         return 0;
 638 }
 639
 640 void kvm_irqfd_exit(void)
 641 {
 642         destroy_workqueue(irqfd_cleanup_wq);
 643 }
 644 #endif
 645
 646 /*
 647  * --------------------------------------------------------------------
 648  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 649  *
 650  * userspace can register a PIO/MMIO address with an eventfd for receiving
 651  * notification when the memory has been touched.
 652  * --------------------------------------------------------------------
 653  */
 654
 655 struct _ioeventfd {
 656         struct list_head     list;
 657         u64                  addr;
 658         int                  length;
 659         struct eventfd_ctx  *eventfd;
 660         u64                  datamatch;
 661         struct kvm_io_device dev;
 662         u8                   bus_idx;
 663         bool                 wildcard;
 664 };
 665
 666 static inline struct _ioeventfd *
 667 to_ioeventfd(struct kvm_io_device *dev)
 668 {
 669         return container_of(dev, struct _ioeventfd, dev);
 670 }
 671
 672 static void
 673 ioeventfd_release(struct _ioeventfd *p)
 674 {
 675         eventfd_ctx_put(p->eventfd);
 676         list_del(&p->list);
 677         kfree(p);
 678 }
 679
 680 static bool
 681 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 682 {
 683         u64 _val;
 684
 685         if (addr != p->addr)
 686                 /* address must be precise for a hit */
 687                 return false;
 688
 689         if (!p->length)
 690                 /* length = 0 means only look at the address, so always a hit */
 691                 return true;
 692
 693         if (len != p->length)
 694                 /* address-range must be precise for a hit */
 695                 return false;
 696
 697         if (p->wildcard)
 698                 /* all else equal, wildcard is always a hit */
 699                 return true;
 700
 701         /* otherwise, we have to actually compare the data */
 702
 703         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 704
 705         switch (len) {
 706         case 1:
 707                 _val = *(u8 *)val;
 708                 break;
 709         case 2:
 710                 _val = *(u16 *)val;
 711                 break;
 712         case 4:
 713                 _val = *(u32 *)val;
 714                 break;
 715         case 8:
 716                 _val = *(u64 *)val;
 717                 break;
 718         default:
 719                 return false;
 720         }
 721
 722         return _val == p->datamatch ? true : false;
 723 }
 724
 725 /* MMIO/PIO writes trigger an event if the addr/val match */
 726 static int
 727 ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
 728                 int len, const void *val)
 729 {
 730         struct _ioeventfd *p = to_ioeventfd(this);
 731
 732         if (!ioeventfd_in_range(p, addr, len, val))
 733                 return -EOPNOTSUPP;
 734
 735         eventfd_signal(p->eventfd, 1);
 736         return 0;
 737 }
 738
 739 /*
 740  * This function is called as KVM is completely shutting down.  We do not
 741  * need to worry about locking just nuke anything we have as quickly as possible
 742  */
 743 static void
 744 ioeventfd_destructor(struct kvm_io_device *this)
 745 {
 746         struct _ioeventfd *p = to_ioeventfd(this);
 747
 748         ioeventfd_release(p);
 749 }
 750
 751 static const struct kvm_io_device_ops ioeventfd_ops = {
 752         .write      = ioeventfd_write,
 753         .destructor = ioeventfd_destructor,
 754 };
 755
 756 /* assumes kvm->slots_lock held */
 757 static bool
 758 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 759 {
 760         struct _ioeventfd *_p;
 761
 762         list_for_each_entry(_p, &kvm->ioeventfds, list)
 763                 if (_p->bus_idx == p->bus_idx &&
 764                     _p->addr == p->addr &&
 765                     (!_p->length || !p->length ||
 766                      (_p->length == p->length &&
 767                       (_p->wildcard || p->wildcard ||
 768                        _p->datamatch == p->datamatch))))
 769                         return true;
 770
 771         return false;
 772 }
 773
 774 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 775 {
 776         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 777                 return KVM_PIO_BUS;
 778         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 779                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 780         return KVM_MMIO_BUS;
 781 }
 782
 783 static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
 784                                 enum kvm_bus bus_idx,
 785                                 struct kvm_ioeventfd *args)
 786 {
 787
 788         struct eventfd_ctx *eventfd;
 789         struct _ioeventfd *p;
 790         int ret;
 791
 792         eventfd = eventfd_ctx_fdget(args->fd);
 793         if (IS_ERR(eventfd))
 794                 return PTR_ERR(eventfd);
 795
 796         p = kzalloc(sizeof(*p), GFP_KERNEL);
 797         if (!p) {
 798                 ret = -ENOMEM;
 799                 goto fail;
 800         }
 801
 802         INIT_LIST_HEAD(&p->list);
 803         p->addr    = args->addr;
 804         p->bus_idx = bus_idx;
 805         p->length  = args->len;
 806         p->eventfd = eventfd;
 807
 808         /* The datamatch feature is optional, otherwise this is a wildcard */
 809         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 810                 p->datamatch = args->datamatch;
 811         else
 812                 p->wildcard = true;
 813
 814         mutex_lock(&kvm->slots_lock);
 815
 816         /* Verify that there isn't a match already */
 817         if (ioeventfd_check_collision(kvm, p)) {
 818                 ret = -EEXIST;
 819                 goto unlock_fail;
 820         }
 821
 822         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 823
 824         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 825                                       &p->dev);
 826         if (ret < 0)
 827                 goto unlock_fail;
 828
 829         kvm->buses[bus_idx]->ioeventfd_count++;
 830         list_add_tail(&p->list, &kvm->ioeventfds);
 831
 832         mutex_unlock(&kvm->slots_lock);
 833
 834         return 0;
 835
 836 unlock_fail:
 837         mutex_unlock(&kvm->slots_lock);
 838
 839 fail:
 840         kfree(p);
 841         eventfd_ctx_put(eventfd);
 842
 843         return ret;
 844 }
 845
 846 static int
 847 kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
 848                            struct kvm_ioeventfd *args)
 849 {
 850         struct _ioeventfd        *p, *tmp;
 851         struct eventfd_ctx       *eventfd;
 852         int                       ret = -ENOENT;
 853
 854         eventfd = eventfd_ctx_fdget(args->fd);
 855         if (IS_ERR(eventfd))
 856                 return PTR_ERR(eventfd);
 857
 858         mutex_lock(&kvm->slots_lock);
 859
 860         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 861                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 862
 863                 if (p->bus_idx != bus_idx ||
 864                     p->eventfd != eventfd  ||
 865                     p->addr != args->addr  ||
 866                     p->length != args->len ||
 867                     p->wildcard != wildcard)
 868                         continue;
 869
 870                 if (!p->wildcard && p->datamatch != args->datamatch)
 871                         continue;
 872
 873                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 874                 kvm->buses[bus_idx]->ioeventfd_count--;
 875                 ioeventfd_release(p);
 876                 ret = 0;
 877                 break;
 878         }
 879
 880         mutex_unlock(&kvm->slots_lock);
 881
 882         eventfd_ctx_put(eventfd);
 883
 884         return ret;
 885 }
 886
 887 static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 888 {
 889         enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
 890         int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 891
 892         if (!args->len && bus_idx == KVM_MMIO_BUS)
 893                 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 894
 895         return ret;
 896 }
 897
 898 static int
 899 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 900 {
 901         enum kvm_bus              bus_idx;
 902         int ret;
 903
 904         bus_idx = ioeventfd_bus_from_flags(args->flags);
 905         /* must be natural-word sized, or 0 to ignore length */
 906         switch (args->len) {
 907         case 0:
 908         case 1:
 909         case 2:
 910         case 4:
 911         case 8:
 912                 break;
 913         default:
 914                 return -EINVAL;
 915         }
 916
 917         /* check for range overflow */
 918         if (args->addr + args->len < args->addr)
 919                 return -EINVAL;
 920
 921         /* check for extra flags that we don't understand */
 922         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 923                 return -EINVAL;
 924
 925         /* ioeventfd with no length can't be combined with DATAMATCH */
 926         if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
 927                 return -EINVAL;
 928
 929         ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
 930         if (ret)
 931                 goto fail;
 932
 933         /* When length is ignored, MMIO is also put on a separate bus, for
 934          * faster lookups.
 935          */
 936         if (!args->len && bus_idx == KVM_MMIO_BUS) {
 937                 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 938                 if (ret < 0)
 939                         goto fast_fail;
 940         }
 941
 942         return 0;
 943
 944 fast_fail:
 945         kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 946 fail:
 947         return ret;
 948 }
 949
 950 int
 951 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 952 {
 953         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 954                 return kvm_deassign_ioeventfd(kvm, args);
 955
 956         return kvm_assign_ioeventfd(kvm, args);
 957 }